Unsupervised anomaly detection method and system for deep-sea manned submersible based on multi-modal feature fusion

Abstract

The present application relates to a deep-sea manned submersible unsupervised anomaly detection method and system based on multi-modal feature fusion, comprising: carrying out initial feature extraction on the pre-processed multi-modal signal data to obtain local time sequence features; carrying out multi-scale frequency domain enhancement on the local time sequence characteristics through a frequency domain enhancement module to obtain frequency domain enhanced features; modeling long-term dependency relationship on the frequency domain enhanced features to obtain long-distance dependency features; using a fusion attention encoding module to carry out context enhancement on the frequency domain enhanced features to obtain encoding features; carrying out multi-modal feature fusion on the frequency domain enhanced features, the long-distance dependency features and the encoding features to obtain fusion features; modeling the fusion features to obtain global hidden state features, and reconstructing and restoring the global hidden state features to obtain reconstruction output; and using a joint loss function to optimize parameters, and through continuous iterative training, obtaining a trained anomaly detection model.

Description

Technical Field

[0001] This invention relates to an unsupervised anomaly detection method and system for deep-sea manned submersibles based on multimodal feature fusion, belonging to the field of artificial intelligence technology. Background Technology

[0002] In high-risk, demanding missions such as scientific experiments, seabed geological structure research, deep-sea shipwreck archaeology, and underwater equipment maintenance, the complexity of deep-sea exploration missions is constantly increasing. Submersible systems are becoming increasingly complex in terms of structure, control, power, sensing, and communication, leading to ever-higher requirements for system safety, reliability, and intelligence. To ensure stable operation of submersibles in extreme environments such as high pressure, strong corrosion, and limited communication in the deep sea, modern deep-sea manned submersibles are typically equipped with numerous sensor modules to monitor key data such as equipment status, environmental parameters, and operational behavior in real time. These sensors encompass various types, including mechanical structure stress sensors, watertight compartment pressure sensors, propulsion system current and voltage monitoring, attitude inertial units (IMUs), depth sensors, water temperature and salinity sensors, and underwater acoustic communication status monitors. During long-term operation, these sensors generate a large amount of high-dimensional, time-series multimodal data, providing a rich source of information for status assessment, fault diagnosis, and safety early warning. However, due to the non-repeatable nature of the deep-sea environment and the irreversible nature of missions, a submersible malfunction during a mission can result in significant economic losses or even endanger personnel lives. Therefore, Fault Detection, Isolation and Health Management (FDIHM) technology has become an important part of submersible design.

[0003] Currently, traditional submersible fault detection mainly relies on rule-based or expert system methods, combining preset thresholds, rule bases, or supervised learning models built based on limited historical samples to determine the state. While these methods perform well under some typical failure modes, they suffer from the following prominent problems: Lack of real-world deep-sea failure samples: Due to the high cost and risk of deep-sea operations, actual failure samples are extremely scarce, making it difficult to collect labeled data covering multiple anomaly modes. This limits the application scope of traditional supervised machine learning algorithms in fault identification; Strong reliance on expert knowledge and experience rules: Traditional model building often requires experienced engineers to manually set fault rules or thresholds, exhibiting strong subjectivity and limitations, making it difficult to generalize or adapt to changing task scenarios; Complex and easily disturbed data modalities: Deep-sea sensor data often has characteristics such as high dimensionality, multimodality, high noise, and non-stationary distribution. Traditional feature extraction methods struggle to extract features with stable discriminative capabilities, affecting detection accuracy; Real-time and deployment challenges: During task execution, the anomaly detection system needs strong real-time response and edge deployment capabilities to operate independently under the limited computing resources of the submersible. Traditional methods often fail to meet these requirements.

[0004] In recent years, with the development of deep learning and artificial intelligence technologies, unsupervised anomaly detection methods have attracted widespread attention due to their ability to handle complex data structures without requiring large amounts of labeled data. Particularly in safety-critical fields such as the Industrial Internet of Things, aerospace, and medical diagnostics, unsupervised methods have gradually demonstrated superior performance in environments with small samples, weak labels, or no labels. In deep-sea manned submersible applications, unsupervised anomaly detection methods have particular application potential; the basic idea is to establish a "benchmark model" of normal behavior by learning the sensor data distribution patterns of the submersible system under normal operating conditions, thereby identifying anomalous data points that deviate from the normal distribution during operation, achieving early warning and fault identification. Typical unsupervised anomaly detection techniques include, but are not limited to: Autoencoders: which learn low-dimensional embeddings and reconstruction errors in data and consider samples with poor reconstruction capabilities as potential anomalies; Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs): which use probabilistic generative models to fit normal data distributions and identify low-probability samples as anomalies; Distance- or density-based methods: such as LOF (Local Outlier Factor), KNN, and Isolation Forest; Temporal modeling methods: such as prediction models based on LSTM and Transformer, which use them to identify anomalies in prediction errors of time series; Graph Neural Networks (GNNs): which are used to model the topological structure and collaborative relationships between multiple sensors; and Multimodal fusion methods: which improve the generalization and robustness of models by fusing data from different types of sensors.

[0005] While the aforementioned unsupervised detection methods have achieved good results on some industrial datasets, several challenges remain in their application to deep-sea manned submersibles. These include: difficulties in multimodal data fusion (due to the diverse types of sensors on deep-sea submersibles and variations in data heterogeneity and latency, making the construction of an effective multimodal feature fusion mechanism a key issue); complex distribution of normal operating states (due to the diverse missions of submersibles and their significant influence from the marine environment, resulting in significant non-stationarity in the distribution of normal data and insufficient generalization ability of benchmark models); and deployment feasibility (due to limitations in submersible computing resources, bandwidth, and energy consumption, traditional large models are difficult to deploy directly in real-world systems, necessitating the design of lightweight, deployable unsupervised detection algorithms).

[0006] In summary, unsupervised anomaly detection for deep-sea manned submersibles still faces multiple challenges such as data scarcity, mission complexity, and computational limitations. There is an urgent need for an unsupervised anomaly detection technology solution that combines multimodal adaptability, high robustness, and deployment feasibility to achieve intelligent monitoring, early warning, and mission safety assurance of deep-sea manned submersibles. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an unsupervised anomaly detection method and system for deep-sea manned submersibles based on multimodal feature fusion. This aims to solve technical problems in existing technologies such as scarcity of anomaly samples, complex data modalities, difficulty in extracting deep temporal patterns, and difficulty in model deployment. The method constructs a fusion-based multi-module anomaly detection model based on deep neural networks, comprehensively utilizing multi-dimensional information such as reconstruction error, attention mechanisms, modal redundancy, and semantic consistency for joint modeling, effectively improving the perception and discrimination capabilities of potential anomalies.

[0008] The technical solution of the present invention is as follows:

[0009] The first aspect of this invention provides an unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion, comprising:

[0010] Multimodal signal data of a deep-sea submersible is collected and preprocessed; an anomaly detection model is constructed, and the preprocessed multimodal signal data is input into the anomaly detection model for training to obtain a trained anomaly detection model. The multimodal signal data of the deep-sea submersible is then input into the trained anomaly detection model to detect anomalies in the deep-sea submersible.

[0011] An anomaly detection model is constructed by inputting preprocessed multimodal signal data into the model for training, including:

[0012] Step 1: Perform initial feature extraction on the preprocessed multimodal signal data to obtain local temporal features;

[0013] Step 2: Multi-scale frequency domain enhancement of local time-series characteristics is performed using the frequency domain enhancement module to obtain frequency domain enhanced features;

[0014] The frequency domain enhancement module (FEM) includes multi-branch convolution operations and fusion operations. The multi-branch convolution operation extracts hidden frequency domain perturbation features, and the fusion operation fuses the frequency domain perturbation features.

[0015] Step 3: Model the long-term dependency relationship of the frequency domain enhancement features to obtain long-distance dependency features;

[0016] Step 4: Use the fusion-based attention coding module to perform contextual enhancement on the frequency domain enhanced features to obtain the coded features;

[0017] The fusion attention encoding module models the correlation between features through serial attention and prior attention, enhancing the ability to capture local dependencies;

[0018] Step 5: Fuse the frequency domain enhancement features, long-range dependency features, and coding features into multimodal features to obtain fused features;

[0019] Step 6: Perform state modeling on the fused features to obtain global hidden state features, and reconstruct and restore the global hidden state features to obtain the reconstruction output;

[0020] Step 7: Optimize the parameters using a joint loss function, and obtain a well-trained anomaly detection model through continuous iterative training.

[0021] According to a preferred embodiment of the present invention, the acquisition and preprocessing of multimodal signal data from a deep-sea submersible includes:

[0022] The data collected from the deep-sea submersible under various operating environments includes multimodal sensor signals, including time-domain signals and frequency-domain signals. The time-domain signals include depth, temperature, attitude signals, acceleration, and angular velocity, while the frequency-domain signals include acoustic positioning signals, pressure pulsation signals, noise signals, and angular vibration signals.

[0023] The multimodal sensor signal data is cleaned, including missing value processing and outlier processing, to generate the input signal. Where B is the batch size, T is the time window length, and C represents the number of channels;

[0024] Further preferably, the preprocessed multimodal signal data undergoes initial feature extraction to obtain local temporal features; including:

[0025] The input signal is used for initial feature extraction to obtain local temporal features, as shown below:

[0026] ;

[0027] in, Represents local temporal features, ReLU represents the activation function, and Conv1D represents a one-dimensional convolutional layer operation, defined as:

[0028] ;

[0029] in, Here are the convolution kernel parameters, K is the kernel size, and C represents the number of channels. The input signal is 'b', where 'b' represents the batch index and 't' represents the current time step index, i.e., the 'k'th time position in the output. This represents the index of the input channel, and k represents the index of the convolution kernel in the time dimension, i.e., the k-th point within the time window. For bias, output The convolutional layer captures local dependency features at adjacent time points, while improving parameter efficiency through weight sharing.

[0030] According to a preferred embodiment of the present invention, multi-scale frequency domain enhancement of local time-series characteristics is performed by a frequency domain enhancement module to obtain frequency domain enhanced features; including:

[0031] The frequency domain enhancement module aims to uncover the hidden frequency domain perturbation features and multi-scale periodic structure information in the original sensor signals, solving the problem of insufficient sensitivity of traditional time series modeling to frequency dependence, thereby improving the anomaly detection system's ability to identify hidden periodic anomalies; this module can capture frequency-related anomaly patterns that are difficult to model using traditional time series methods, and improve detection sensitivity by adaptively capturing frequency domain anomaly features through learning convolution kernel weights.

[0032] The multi-branch convolution operation consists of two branches. The obtained local temporal features are input into the two branches respectively, and convolution, normalization, and activation function processing are performed sequentially to obtain frequency domain perturbation features, as shown below:

[0033] ;

[0034] ;

[0035] in, and This represents the two frequency domain perturbation characteristics obtained. Representing local temporal features B is the batch size, C is the number of channels, and T is the time series length. Indicates batch normalization. This represents a one-dimensional convolution with kernel K. This represents dilated convolution, where K is the kernel size (K=3) and d is the kernel sampling interval (d=2).

[0036] The two frequency domain perturbation features are fused using a fusion operation to obtain the frequency domain enhanced features:

[0037] ;

[0038] ;

[0039] in, Indicates frequency domain enhancement features, Indicates splicing characteristics, This indicates a splicing operation. This represents a one-dimensional convolution with kernel K1. This indicates the kernel size (K1=1). Indicates the number of output channels.

[0040] According to a preferred embodiment of the present invention, long-term dependency modeling is performed on the frequency domain enhancement features to obtain long-distance dependency features; including:

[0041] The Mamba module is introduced to enhance the internal features of sequences and perform long-term dependency modeling. As an emerging structure with higher sequence modeling efficiency than Transformer, Mamba can effectively make up for the performance degradation of traditional Transformer for long sequence modeling. The Mamba module can effectively integrate the temporal correlation and cross-modal semantic dependency between different modalities, and enhance the robustness and generalization ability of the model.

[0042] The Mamba module includes splicing and merging operations, convolution operations, state space operations, and dynamic gating operations;

[0043] The frequency domain enhancement features are divided into M feature sequences, and then concatenated and fused as follows:

[0044] ;

[0045] in, This indicates the characteristics of splicing and merging. This represents the feedforward network for the Mth feature sequence. Let M represent the Mth feature sequence, where M represents the number of feature sequences for frequency domain enhancement features;

[0046] Calculate the mapped input:

[0047] ;

[0048] in, Indicates the mapped input, Represents the weight matrix. The dimension of the feature vector at each time step is represented by ⊙, where ⊙ denotes element-wise multiplication. It is the Sigmoid activation function. For activation function, and Represents the paranoia vector;

[0049] Use convolution operations to map the input. Further local features are extracted, and nonlinear mapping is performed using anomaly-aware GELU:

[0050] = ;

[0051] At the same time, an anomalous enhancement activation mechanism is applied to the mapped input, as shown below:

[0052] = ;

[0053] in, Represents mapping features, Indicates enhanced features, and Conv represents the convolution operation. Indicates the activation function;

[0054] Using state-space operations (selective SSM) to process long sequence data with mapped features:

[0055] ;

[0056] in, This represents the hidden state at time t. This represents the hidden state at time t-1. A represents the mapping characteristics at time t. For trainable parameters, sliding window convolution optimization is used to accelerate computation and support long sequence modeling;

[0057] High-response channels are extracted through dynamic gating operations to obtain long-distance dependency features:

[0058] ;

[0059] in, This represents the activation function. Represents the weight vector. This represents the bias vector. Represents the hidden state sequence, This indicates long-distance dependency features.

[0060] According to a preferred embodiment of the present invention, a fusion attention coding module is used to perform contextual enhancement on the frequency domain enhanced features to obtain coded features; including:

[0061] The fusion-based attention encoding module includes encoding operations and anomaly attention operations.

[0062] Abnormal attention operations include serial attention and prior attention;

[0063] A fusion-based attention encoding module is used to enhance the context of long-distance dependent features. The fusion-based attention encoding module (Anomaly Transformer Encoder) adopts an improved Transformer structure and embeds an AnomalyAttention mechanism. It models the correlation between normal behavior and potential abnormal behavior based on local attention within a time window. The output of the fusion-based attention encoding module includes the encoding result, attention series information, prior attention distribution and variance information, providing a basis for subsequent anomaly judgment.

[0064] The frequency domain enhancement features are encoded as follows:

[0065] ;

[0066] in, Indicates the input encoding features, Indicates word embedding;

[0067] Calculate the query vector, value vector, and key vector of the encoded result:

[0068] , ;

[0069] in, Represents the query vector. V represents the key vector, and V represents the value vector. , , Represents the weight matrix;

[0070] Computational series attention characteristics:

[0071] ;

[0072] in, Represents a series of attentional features, Represents the dimension of the key vector G;

[0073] Calculate prior attention features:

[0074] ;

[0075] in, This represents the prior attention matrix, i.e., the prior attention features. The variance is represented by i and j, which represent the positions of two time steps, respectively. The prior matrix introduces prior knowledge based on positional relationships to enhance the model's ability to capture local dependencies.

[0076] The calculated prior attention matrix With a series of attentional characteristics The final attention weight matrix is ​​generated by fusion. As shown below:

[0077] ;

[0078] The calculated encoding result is shown below:

[0079] ;

[0080] in, These are encoding features.

[0081] According to a preferred embodiment of the present invention, frequency domain enhancement features, long-range dependency features, and coding features are fused using multimodal feature fusion to obtain fused features; including:

[0082] The multimodal feature fusion module (WeightedFeatureFusion) is used to perform weighted fusion of frequency domain enhanced features, long-distance dependency features and coded features, and the fusion weight parameters are obtained through training. Compared with the traditional simple concatenation method, this module can achieve more semantically consistent and feature-complementary effective fusion, which helps to improve the stability and robustness of the model.

[0083] Set the frequency domain enhancement feature as Long-distance dependency features are set as The encoding features are set as follows: ;

[0084] Calculate the global weight for each feature:

[0085] ;

[0086] in, represents the global weight of the i-th feature; t represents the time step; Let i represent the i-th feature among the three types of features, and T be the length of the time window;

[0087] Calculate the weights of the linear mapping:

[0088] ;

[0089] in, This represents the linear mapping result of the i-th type of feature. Represents the weight matrix. Indicates the bias term;

[0090] Normalization is performed:

[0091] ;

[0092] in, Represents the normalized weights of the i-th type of feature. This represents the linear mapping result of the j-th type of feature;

[0093] Finally, the fusion features are obtained:

[0094] ;

[0095] in, This indicates the fusion feature.

[0096] According to a preferred embodiment of the present invention, state modeling is performed on the fused features to obtain global hidden state features, and the global hidden state features are reconstructed and restored to obtain a reconstructed output; including:

[0097] Feature extraction is performed on the fused features as follows:

[0098] ;

[0099] in, This represents the hidden state at the current time step t. This represents the hidden state at time step t−1. This indicates bidirectional GRU network (gated cyclic unit) operation;

[0100] By compressing the time dimension and obtaining the final state representation, we obtain the global hidden state features:

[0101] ;

[0102] in, Represents the global hidden state features. For the hidden layer dimension, Indicates the last time step The hidden state;

[0103] The global hidden state features are reconstructed and restored. The reconstruction and restoration calculation process is as follows:

[0104] ;

[0105] in, Indicates reconstruction output. Indicates a decoding operation;

[0106] The error between the input signal X and the reconstructed output is calculated as follows:

[0107] ;

[0108] in, Indicates reconstruction error. Indicates the input signal. This represents the reconstructed output, where B, T, and C are the batch size, time window length, and number of channels, respectively. These represent the current batch size, the current time window length, and the current number of channels, respectively.

[0109] If there is a significant deviation between the original input (i.e., the input signal) and the reconstructed output, i.e., the reconstruction error is greater than the set threshold ( If the input signal is not found, the input signal is considered an abnormal signal and the output detection result is abnormal; otherwise, it is a normal signal.

[0110] A further preferred approach involves using a joint loss function for parameter optimization, followed by iterative training to obtain a well-trained anomaly detection model; including:

[0111] Constructing Reconstruction Loss :

[0112] ;

[0113] in, Indicates the input signal. Indicates reconstruction output;

[0114] Constructing a normalized prior distribution :

[0115] ;

[0116] in, This is the normalized prior distribution. Indicates the total length of the sequence. This represents the original prior distribution value, i.e., the initial probability distribution, at time step v. This represents the original prior distribution value at time step u;

[0117] Construct a series loss:

[0118] ;

[0119] in, Indicates a series of consistency losses. This is the normalized prior distribution. Indicates the Kullback-Leibler divergence. This represents the model prediction distribution at time step u (the probabilistic prediction of sequence features made by the model at time step u based on the current input signal and previously learned knowledge).

[0120] Construct the prior stability loss:

[0121] ;

[0122] in, This represents the prior stability loss. Indicates to Smooth and enhance contrast. This represents the original prior distribution value at time step u;

[0123] Define two loss functions with relative objectives: minimizing the objective and maximizing the objective. The minimizing objective (used to update the generator network) is:

[0124] ;

[0125] The objective to maximize (for training on a stable distribution) is:

[0126] ;

[0127] in, To adjust the parameters;

[0128] During training, by minimizing the target loss and maximizing target loss Optimize and iterate the training process to obtain a well-trained anomaly detection model.

[0129] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of an unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion.

[0130] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of an unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion.

[0131] A second aspect of the present invention provides an unsupervised anomaly detection system for deep-sea manned submersibles based on multimodal feature fusion, comprising:

[0132] The model building module is configured to: collect multimodal signal data of deep-sea submersibles and preprocess it; build an anomaly detection model, input the preprocessed multimodal signal data into the anomaly detection model for training, obtain a trained anomaly detection model, and input the multimodal signal data of deep-sea submersibles into the trained anomaly detection model for anomaly detection of deep-sea submersibles;

[0133] The initial feature extraction module is configured to extract initial features from the preprocessed multimodal signal data to obtain local temporal features;

[0134] The frequency domain enhancement module is configured to perform multi-scale frequency domain enhancement on local time-series characteristics to obtain frequency domain enhanced features.

[0135] The frequency domain enhancement module includes multi-branch convolution operations and fusion operations. The multi-branch convolution operation extracts hidden frequency domain perturbation features, and the fusion operation fuses the frequency domain perturbation features.

[0136] The long-term dependency modeling module is configured to: perform long-term dependency modeling on frequency domain enhanced features to obtain long-distance dependency features;

[0137] The context enhancement module is configured to: use the fusion attention coding module to perform context enhancement on the frequency domain enhanced features to obtain coded features;

[0138] The fusion attention encoding module models the correlation between features through serial attention and prior attention, enhancing the ability to capture local dependencies;

[0139] The feature fusion module is configured to fuse frequency domain enhanced features, long-range dependency features, and coding features into multimodal features to obtain fused features;

[0140] The reconstruction module is configured to: perform state modeling on the fused features, obtain global hidden state features, and reconstruct and restore the global hidden state features to obtain the reconstruction output;

[0141] The optimization module is configured to use a joint loss function to optimize parameters and obtain a trained anomaly detection model through continuous iterative training.

[0142] The beneficial effects of this invention are as follows:

[0143] 1. Enhanced detection robustness through multimodal structural fusion: This invention integrates multiple feature extraction mechanisms in the time domain, frequency domain, structural domain, and attention domain, enabling the model to have stronger stability and adaptability when facing high-noise and multi-source heterogeneous data in the marine environment.

[0144] 2. Introducing the Mamba module structure to optimize long sequence modeling capabilities: Compared with the traditional Transformer, Mamba has better sequence modeling capabilities and computational efficiency, making it suitable for processing long-term historical data analysis in submersible missions.

[0145] 3. Reconstruction mechanism achieves zero-sample anomaly detection: By constructing a reconstruction model in a normal state, it does not need to rely on any labeled anomaly samples, thus overcoming the problem of scarce anomaly samples in deep-sea missions and achieving unsupervised anomaly detection.

[0146] 4. Feature fusion learnable weight mechanism: The designed weighted fusion module can adaptively adjust the importance of features from different sources, avoiding subjective bias caused by manually setting fusion strategies.

[0147] 5. End-to-end modeling and deployment friendliness: The overall model supports end-to-end training and inference, and can be tailored into a lightweight version for deployment on edge computing platforms, meeting the deployment requirements of deep-sea missions for real-time performance and resource constraints.

[0148] 6. High interpretability and easy anomaly tracing: The attention map and prior distribution information output by the encoder can be used to analyze the key areas of interest of the model, assisting engineers in anomaly tracing and analysis. Attached Figure Description

[0149] Figure 1 This is a flowchart of the unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion according to the present invention;

[0150] Figure 2 This is a flowchart of the data processing based on projection and selective SSM according to the present invention;

[0151] Figure 3 This is a flowchart of the data processing based on the attention mechanism of the present invention. Detailed Implementation

[0152] The present invention will be further described below with reference to the embodiments and accompanying drawings, but is not limited thereto.

[0153] Example 1

[0154] An unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion, such as Figure 1 As shown, it includes:

[0155] Multimodal signal data of a deep-sea submersible is collected and preprocessed; an anomaly detection model is constructed, and the preprocessed multimodal signal data is input into the anomaly detection model for training to obtain a trained anomaly detection model. The multimodal signal data of the deep-sea submersible is then input into the trained anomaly detection model to detect anomalies in the deep-sea submersible.

[0156] An anomaly detection model is constructed by inputting preprocessed multimodal signal data into the model for training, including:

[0157] Step 1: Perform initial feature extraction on the preprocessed multimodal signal data to obtain local temporal features;

[0158] Step 2: Multi-scale frequency domain enhancement of local time-series characteristics is performed using the frequency domain enhancement module to obtain frequency domain enhanced features;

[0159] The frequency domain enhancement module (FEM) includes multi-branch convolution operations and fusion operations. The multi-branch convolution operation extracts hidden frequency domain perturbation features, and the fusion operation fuses the frequency domain perturbation features.

[0160] Step 3: Model the long-term dependency relationship of the frequency domain enhancement features to obtain long-distance dependency features;

[0161] Step 4: Use the fusion-based attention coding module to perform contextual enhancement on the frequency domain enhanced features to obtain the coded features;

[0162] The fusion attention encoding module models the correlation between features through serial attention and prior attention, enhancing the ability to capture local dependencies;

[0163] Step 5: Fuse the frequency domain enhancement features, long-range dependency features, and coding features into multimodal features to obtain fused features;

[0164] Step 6: Perform state modeling on the fused features to obtain global hidden state features, and reconstruct and restore the global hidden state features to obtain the reconstruction output;

[0165] Step 7: Optimize the parameters using a joint loss function, and obtain a well-trained anomaly detection model through continuous iterative training.

[0166] Example 2

[0167] The difference between the unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion described in Example 1 and the method described in Example 1 is as follows:

[0168] Acquire multimodal signal data from deep-sea submersibles and perform preprocessing; including:

[0169] The data collected from the deep-sea submersible under various operating environments includes multimodal sensor signals, including time-domain signals and frequency-domain signals. The time-domain signals include depth, temperature, attitude signals, acceleration, and angular velocity, while the frequency-domain signals include acoustic positioning signals, pressure pulsation signals, noise signals, and angular vibration signals.

[0170] The multimodal sensor signal data is cleaned, including missing value processing and outlier processing, to generate the input signal. Where B is the batch size, T is the time window length, and C represents the number of channels;

[0171] Initial feature extraction is performed on the preprocessed multimodal signal data to obtain local temporal features, including:

[0172] The input signal is used for initial feature extraction to obtain local temporal features, as shown below:

[0173] ;

[0174] in, Represents local temporal features, ReLU represents the activation function, and Conv1D represents a one-dimensional convolutional layer operation, defined as:

[0175] ;

[0176] in, Here are the convolution kernel parameters, K is the kernel size, and C represents the number of channels. The input signal is 'b', where 'b' represents the batch index and 't' represents the current time step index, i.e., the 'k'th time position in the output. This represents the index of the input channel, and k represents the index of the convolution kernel in the time dimension, i.e., the k-th point within the time window. For bias, output The convolutional layer captures local dependency features at adjacent time points, while improving parameter efficiency through weight sharing.

[0177] The local time-series characteristics are enhanced in a multi-scale frequency domain using a frequency domain enhancement module to obtain enhanced frequency domain features, including:

[0178] The frequency domain enhancement module aims to uncover the hidden frequency domain perturbation features and multi-scale periodic structure information in the original sensor signals, solving the problem of insufficient sensitivity of traditional time series modeling to frequency dependence, thereby improving the anomaly detection system's ability to identify hidden periodic anomalies; this module can capture frequency-related anomaly patterns that are difficult to model using traditional time series methods, and improve detection sensitivity by adaptively capturing frequency domain anomaly features through learning convolution kernel weights.

[0179] The multi-branch convolution operation consists of two branches. The obtained local temporal features are input into the two branches respectively, and convolution, normalization, and activation function processing are performed sequentially to obtain frequency domain perturbation features, as shown below:

[0180] ;

[0181] ;

[0182] in, and This represents the two frequency domain perturbation characteristics obtained. Representing local temporal features B is the batch size, C is the number of channels, and T is the time series length. Indicates batch normalization. This represents a one-dimensional convolution with kernel K. This represents dilated convolution, where K is the kernel size (K=3) and d is the kernel sampling interval (d=2).

[0183] The two frequency domain perturbation features are fused using a fusion operation to obtain the frequency domain enhanced features:

[0184] ;

[0185] ;

[0186] in, Indicates frequency domain enhancement features, Indicates splicing characteristics, This indicates a splicing operation. This represents a one-dimensional convolution with kernel K1. This indicates the kernel size (K1=1). Indicates the number of output channels.

[0187] Long-term dependency modeling is performed on frequency domain enhancement features to obtain long-distance dependency features; such as Figure 2 As shown, it includes:

[0188] The Mamba module is introduced to enhance the internal features of sequences and perform long-term dependency modeling. As an emerging structure with higher sequence modeling efficiency than Transformer, Mamba can effectively make up for the performance degradation of traditional Transformer for long sequence modeling. The Mamba module can effectively integrate the temporal correlation and cross-modal semantic dependency between different modalities, and enhance the robustness and generalization ability of the model.

[0189] The Mamba module includes splicing and merging operations, convolution operations, state space operations, and dynamic gating operations;

[0190] The frequency domain enhancement features are divided into M feature sequences, and then concatenated and fused as follows:

[0191] ;

[0192] in, This indicates the characteristics of splicing and merging. This represents the feedforward network for the Mth feature sequence. Let M represent the Mth feature sequence, where M represents the number of feature sequences for frequency domain enhancement features;

[0193] Calculate the mapped input:

[0194] ;

[0195] in, Indicates the mapped input, Represents the weight matrix. The dimension of the feature vector at each time step is represented by ⊙, where ⊙ denotes element-wise multiplication. It is the Sigmoid activation function. For activation function, and Represents the paranoia vector;

[0196] Use convolution operations to map the input. Further local features are extracted, and nonlinear mapping is performed using anomaly-aware GELU:

[0197] = ;

[0198] At the same time, an anomalous enhancement activation mechanism is applied to the mapped input, as shown below:

[0199] = ;

[0200] in, Represents mapping features, Indicates enhanced features, and Conv represents the convolution operation. Indicates the activation function;

[0201] Use state-space operations (i.e., selective SSM) to process long sequence data with mapped features:

[0202] ;

[0203] in, This represents the hidden state at time t. This represents the hidden state at time t-1. A represents the mapping characteristics at time t. For trainable parameters, sliding window convolution optimization is used to accelerate computation and support long sequence modeling;

[0204] High-response channels are extracted through dynamic gating operations to obtain long-distance dependency features:

[0205] ;

[0206] in, This represents the activation function. Represents the weight vector. This represents the bias vector. Represents the hidden state sequence, This indicates long-distance dependency features.

[0207] A fusion-based attention coding module is used to enhance the context of the frequency domain enhanced features, resulting in coded features; for example... Figure 3 As shown, it includes:

[0208] The fusion-based attention encoding module includes encoding operations and anomaly attention operations.

[0209] Abnormal attention operations include serial attention and prior attention;

[0210] A fusion-based attention encoding module is used to enhance the context of long-distance dependent features. The fusion-based attention encoding module (Anomaly Transformer Encoder) adopts an improved Transformer structure and embeds an AnomalyAttention mechanism. It models the correlation between normal behavior and potential abnormal behavior based on local attention within a time window. The output of the fusion-based attention encoding module includes the encoding result, attention series information, prior attention distribution and variance information, providing a basis for subsequent anomaly judgment.

[0211] The frequency domain enhancement features are encoded as follows:

[0212] ;

[0213] in, Indicates the input encoding features, Indicates word embedding;

[0214] Calculate the query vector, value vector, and key vector of the encoded result:

[0215] , ;

[0216] in, Represents the query vector. V represents the key vector, and V represents the value vector. , , Represents the weight matrix;

[0217] Computational series attention characteristics:

[0218] ;

[0219] in, Represents a series of attentional features, Represents the dimension of the key vector G;

[0220] Calculate prior attention features:

[0221] ;

[0222] in, This represents the prior attention matrix, i.e., the prior attention features. The variance is represented by i and j, which represent the positions of two time steps, respectively. The prior matrix introduces prior knowledge based on positional relationships to enhance the model's ability to capture local dependencies.

[0223] The calculated prior attention matrix With a series of attentional characteristics The final attention weight matrix is ​​generated by fusion. As shown below:

[0224] ;

[0225] The calculated encoding result is shown below:

[0226] ;

[0227] in, These are encoding features.

[0228] Multimodal feature fusion is performed on frequency domain enhancement features, long-range dependency features, and coding features to obtain fused features, including:

[0229] The multimodal feature fusion module (WeightedFeatureFusion) is used to perform weighted fusion of frequency domain enhanced features, long-distance dependency features and coded features, and the fusion weight parameters are obtained through training. Compared with the traditional simple concatenation method, this module can achieve more semantically consistent and feature-complementary effective fusion, which helps to improve the stability and robustness of the model.

[0230] Set the frequency domain enhancement feature as Long-distance dependency features are set as The encoding features are set as follows: ;

[0231] Calculate the global weight for each feature:

[0232] ;

[0233] in, represents the global weight of the i-th feature; t represents the time step; Let i represent the i-th feature among the three types of features, and T be the length of the time window;

[0234] Calculate the weights of the linear mapping:

[0235] ;

[0236] in, This represents the linear mapping result of the i-th type of feature. Represents the weight matrix. Indicates the bias term;

[0237] Normalization is performed:

[0238] ;

[0239] in, Represents the normalized weights of the i-th type of feature. This represents the linear mapping result of the j-th type of feature;

[0240] Finally, the fusion features are obtained:

[0241] ;

[0242] in, This indicates the fusion feature.

[0243] State modeling is performed on the fused features to obtain global hidden state features, and the global hidden state features are reconstructed and restored to obtain the reconstruction output; including:

[0244] Feature extraction is performed on the fused features as follows:

[0245] ;

[0246] in, This represents the hidden state at the current time step t. This represents the hidden state at time step t−1. This indicates bidirectional GRU network (gated cyclic unit) operation;

[0247] By compressing the time dimension and obtaining the final state representation, we obtain the global hidden state features:

[0248] ;

[0249] in, Represents the global hidden state features. For the hidden layer dimension, Indicates the last time step The hidden state;

[0250] The global hidden state features are reconstructed and restored. The reconstruction and restoration calculation process is as follows:

[0251] ;

[0252] in, Indicates reconstruction output. Indicates a decoding operation;

[0253] The error between the input signal X and the reconstructed output is calculated as follows:

[0254] ;

[0255] in, Indicates reconstruction error. Indicates the input signal. This represents the reconstructed output, where B, T, and C are the batch size, time window length, and number of channels, respectively. These represent the current batch size, the current time window length, and the current number of channels, respectively.

[0256] If there is a significant deviation between the original input (i.e., the input signal) and the reconstructed output, i.e., the reconstruction error is greater than the set threshold ( If the input signal is not found, the input signal is considered an abnormal signal and the output detection result is abnormal; otherwise, it is a normal signal.

[0257] A joint loss function is used for parameter optimization, and a well-trained anomaly detection model is obtained through continuous iterative training; including:

[0258] Constructing Reconstruction Loss :

[0259] ;

[0260] in, Indicates the input signal. Indicates reconstruction output;

[0261] Constructing a normalized prior distribution :

[0262] ;

[0263] in, This is the normalized prior distribution. Indicates the total length of the sequence. This represents the original prior distribution value, i.e., the initial probability distribution, at time step v. This represents the original prior distribution value at time step u;

[0264] Construct a series loss:

[0265] ;

[0266] in, Indicates a series of consistency losses. This is the normalized prior distribution. Indicates the Kullback-Leibler divergence. This represents the model prediction distribution at time step u (the probabilistic prediction of sequence features made by the model at time step u based on the current input signal and previously learned knowledge).

[0267] Construct the prior stability loss:

[0268] ;

[0269] in, This represents the prior stability loss. Indicates to Smooth and enhance contrast. This represents the original prior distribution value at time step u;

[0270] Define two loss functions with relative objectives: minimizing the objective and maximizing the objective. The minimizing objective (used to update the generator network) is:

[0271] ;

[0272] The objective to maximize (for training on a stable distribution) is:

[0273] ;

[0274] in, To adjust the parameters;

[0275] During training, by minimizing the target loss and maximizing target loss Optimize and iterate the training process to obtain a well-trained anomaly detection model.

[0276] The Mamba module is a new (LLM) architecture that integrates a structured state model to manage lengthy data sequences. It combines the best features of both LLM and continuous-time models to effectively model long-term dependencies.

[0277] Bidirectional GRU network: A recurrent neural network that combines gated recurrent units (GRUs) with a bidirectional structure. It captures the past and future contextual information of the sequence at each time step through two GRU layers, forward and backward. It controls the information flow using update and reset gates, effectively modeling long-term dependencies and is suitable for tasks requiring contextual understanding, such as speech recognition and text annotation.

[0278] The effectiveness of the invention was verified on the SWAT, OWN-1, and OWN-2 datasets. SWAT is a publicly available dataset that records various operations and events occurring in the industrial control system of a water treatment plant. These records can include sensor readings, control commands, equipment status changes, etc. OWN-1 and OWN-2 are data collected from different latency levels. The performance of the anomaly detection method in this invention was evaluated using Accuracy, Precision, Recall, and F-score as evaluation metrics, as shown in Table 1.

[0279] Table 1. Anomaly detection performance on different datasets

[0280]

[0281] The proposed method achieved high accuracy, precision, and F-score (0.9926, 0.9711, and 0.9854) on the OWN-1 dataset; and high recall (1.0000) on both the OWN-1 and OWN-2 datasets. The experimental results demonstrate the effectiveness and robustness of the proposed method, which exhibits good adaptability and the ability to detect unknown anomalies.

[0282] Example 3

[0283] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion as described in Embodiment 1 or 2.

[0284] Example 4

[0285] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion as described in Embodiment 1 or 2.

[0286] Example 5

[0287] An unsupervised anomaly detection system for deep-sea manned submersibles based on multimodal feature fusion includes:

[0288] The model building module is configured to: collect multimodal signal data of deep-sea submersibles and preprocess it; build an anomaly detection model, input the preprocessed multimodal signal data into the anomaly detection model for training, obtain a trained anomaly detection model, and input the multimodal signal data of deep-sea submersibles into the trained anomaly detection model for anomaly detection of deep-sea submersibles;

[0289] The initial feature extraction module is configured to extract initial features from the preprocessed multimodal signal data to obtain local temporal features;

[0290] The frequency domain enhancement module is configured to perform multi-scale frequency domain enhancement on local time-series characteristics to obtain frequency domain enhanced features.

[0291] The frequency domain enhancement module includes multi-branch convolution operations and fusion operations. The multi-branch convolution operation extracts hidden frequency domain perturbation features, and the fusion operation fuses the frequency domain perturbation features.

[0292] The long-term dependency modeling module is configured to: perform long-term dependency modeling on frequency domain enhanced features to obtain long-distance dependency features;

[0293] The context enhancement module is configured to: use the fusion attention coding module to perform context enhancement on the frequency domain enhanced features to obtain coded features;

[0294] The fusion attention encoding module models the correlation between features through serial attention and prior attention, enhancing the ability to capture local dependencies;

[0295] The feature fusion module is configured to fuse frequency domain enhanced features, long-range dependency features, and coding features into multimodal features to obtain fused features;

[0296] The reconstruction module is configured to: perform state modeling on the fused features, obtain global hidden state features, and reconstruct and restore the global hidden state features to obtain the reconstruction output;

[0297] The optimization module is configured to use a joint loss function to optimize parameters and obtain a trained anomaly detection model through continuous iterative training.

Claims

1. An unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion, characterized in that, include: Collect multimodal signal data from deep-sea submersibles and perform preprocessing; An anomaly detection model is constructed. The preprocessed multimodal signal data is input into the anomaly detection model for training to obtain a trained anomaly detection model. The multimodal signal data of the deep-sea submersible is then input into the trained anomaly detection model to detect anomalies in the deep-sea submersible. An anomaly detection model is constructed by inputting preprocessed multimodal signal data into the model for training, including: Step 1: Perform initial feature extraction on the preprocessed multimodal signal data to obtain local temporal features; Step 2: Multi-scale frequency domain enhancement of local time-series characteristics is performed using the frequency domain enhancement module to obtain frequency domain enhanced features; The frequency domain enhancement module (FEM) includes multi-branch convolution operations and fusion operations. The multi-branch convolution operation extracts hidden frequency domain perturbation features, and the fusion operation fuses the frequency domain perturbation features. Step 3: Model the long-term dependency relationship of the frequency domain enhancement features to obtain long-distance dependency features; Step 4: Use the fusion-based attention coding module to perform contextual enhancement on the frequency domain enhanced features to obtain the coded features; The fusion attention encoding module models the correlation between features through serial attention and prior attention, enhancing the ability to capture local dependencies; Step 5: Fuse the frequency domain enhancement features, long-range dependency features, and coding features into multimodal features to obtain fused features; Step 6: Perform state modeling on the fused features to obtain global hidden state features, and reconstruct and restore the global hidden state features to obtain the reconstruction output; Step 7: Optimize the parameters using a joint loss function, and obtain a well-trained anomaly detection model through continuous iterative training.

2. The unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion according to claim 1, characterized in that, Acquire multimodal signal data from deep-sea submersibles and perform preprocessing; including: The data collected from the deep-sea submersible under various operating environments includes multimodal sensor signals, including time-domain signals and frequency-domain signals. The time-domain signals include depth, temperature, attitude signals, acceleration, and angular velocity, while the frequency-domain signals include acoustic positioning signals, pressure pulsation signals, noise signals, and angular vibration signals. The multimodal sensor signal data is cleaned, including missing value processing and outlier processing, to generate the input signal. Where B is the batch size, T is the time window length, and C represents the number of channels; Further preferably, the preprocessed multimodal signal data undergoes initial feature extraction to obtain local temporal features; including: The input signal is used for initial feature extraction to obtain local temporal features, as shown below: ； in, Represents local temporal features, ReLU represents the activation function, and Conv1D represents a one-dimensional convolutional layer operation.

3. The unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion according to claim 2, characterized in that, The local time-series characteristics are enhanced in a multi-scale frequency domain using a frequency domain enhancement module to obtain enhanced frequency domain features, including: The multi-branch convolution operation consists of two branches. The obtained local temporal features are input into the two branches respectively, and convolution, normalization, and activation function processing are performed sequentially to obtain frequency domain perturbation features, as shown below: ； ； in, and This represents the two frequency domain perturbation characteristics obtained. Representing local temporal features B is the batch size, C is the number of channels, and T is the time series length. Indicates batch normalization. Represents one-dimensional convolution. This represents dilated convolution, where K is the kernel size and d is the kernel sampling interval. The two frequency domain perturbation features are fused using a fusion operation to obtain the frequency domain enhanced features: ； ； in, Indicates frequency domain enhancement features, Indicates splicing characteristics, This indicates a splicing operation. Represents one-dimensional convolution. Indicates the kernel size. Indicates the number of output channels.

4. The unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion according to claim 3, characterized in that, Long-term dependency modeling is performed on frequency domain enhancement features to obtain long-distance dependency features; including: The Mamba module is introduced to enhance the internal features of sequences and to model long-term dependencies. The Mamba module includes splicing and merging operations, convolution operations, state space operations, and dynamic gating operations; The frequency domain enhancement features are divided into M feature sequences, and then concatenated and fused as follows: ； in, This indicates the characteristics of splicing and merging. This represents a feedforward network for the Mth feature sequence. Let M represent the Mth feature sequence, where M represents the number of feature sequences for frequency domain enhancement features; Calculate the mapped input: ； in, Indicates the mapped input, Represents the weight matrix. The dimension of the feature vector at each time step is represented by ⊙, where ⊙ denotes element-wise multiplication. It is the Sigmoid activation function. For activation function, and Represents the paranoia vector; Use convolution operations to map the input. Further extract local features and perform nonlinear mapping through an anomaly-enhanced activation mechanism: = ； At the same time, an anomalous enhancement activation mechanism is applied to the mapped input, as shown below: = ； in, Represents mapping features, Indicates enhanced features, and Conv represents the convolution operation. Indicates the activation function; Using state-space operations to process long sequence data with mapped features: ； in, This represents the hidden state at time t. This represents the hidden state at time t-1. A represents the mapping characteristics at time t. These are trainable parameters; High-response channels are extracted through dynamic gating operations to obtain long-distance dependency features: ； in, This represents the activation function. Represents the weight vector. This represents the bias vector. Represents the hidden state sequence, This indicates long-distance dependency features.

5. The unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion according to claim 4, characterized in that, The frequency domain enhanced features are enhanced with context using a fusion attention coding module to obtain coded features, including: The fusion-based attention encoding module includes encoding operations and anomaly attention operations; Abnormal attention operations include serial attention and prior attention; The frequency domain enhancement features are encoded as follows: ； in, Indicates the input encoding features, Indicates word embedding; Calculate the query vector, value vector, and key vector of the encoded result: ， ； in, Represents the query vector. V represents the key vector, and V represents the value vector. , , Represents the weight matrix; Computational series attention characteristics: ； in, Represents a series of attentional features, Represents the dimension of the key vector G; Calculate prior attention features: ； in, This represents the prior attention matrix, i.e., the prior attention features. Let i be the variance, and j represent the positions of the two time steps, respectively. The calculated prior attention matrix With a series of attentional characteristics The final attention weight matrix is ​​generated by fusion. As shown below: ； The calculated encoding result is shown below: ； in, These are encoding features.

6. The unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion according to claim 5, characterized in that, Multimodal feature fusion is performed on frequency domain enhancement features, long-range dependency features, and coding features to obtain fused features, including: The multimodal feature fusion module is used to perform weighted fusion of frequency domain enhanced features, long-distance dependency features and coding features, and the fusion weight parameters are obtained through training; Set the frequency domain enhancement feature as Long-distance dependency features are set as The encoding features are set as follows: ; Calculate the global weight for each feature: ； in, represents the global weight of the i-th feature; t represents the time step; Let i represent the i-th feature among the three types of features, and T be the length of the time window; Calculate the weights of the linear mapping: ； in, This represents the linear mapping result of the i-th type of feature. Represents the weight matrix. Indicates the bias term; Normalization is performed: ； in, Represents the normalized weight of the i-th feature. This represents the linear mapping result of the j-th type of feature; Finally, the fusion features are obtained: ； in, This indicates the fusion feature.

7. The unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion according to claim 6, characterized in that, State modeling is performed on the fused features to obtain global hidden state features, and the global hidden state features are reconstructed and restored to obtain the reconstruction output; include: Feature extraction is performed on the fused features as follows: ； in, This represents the hidden state at the current time step t. This represents the hidden state at time step t−1. Indicates a bidirectional GRU network; By compressing the time dimension and obtaining the final state representation, we obtain the global hidden state features: ； in, Represents the global hidden state features. For the hidden layer dimension, Indicates the last time step The hidden state; The global hidden state features are reconstructed and restored. The reconstruction and restoration calculation process is as follows: ； in, Indicates reconstruction output. Indicates a decoding operation; The error between the input signal X and the reconstructed output is calculated as follows: ； in, Indicates reconstruction error. Indicates the input signal. This represents the reconstructed output, where B, T, and C are the batch size, time window length, and number of channels, respectively. These represent the current batch size, the current time window length, and the current number of channels, respectively. If there is a significant deviation between the original input (i.e., the input signal) and the reconstructed output (i.e., the reconstruction error is greater than the set threshold), the input signal is considered an abnormal signal and the output detection result is abnormal; otherwise, it is a normal signal. Furthermore, a joint loss function is used for parameter optimization, and through continuous iterative training, a well-trained anomaly detection model is obtained; including: Constructing Reconstruction Loss : ； in, Indicates the input signal. Indicates reconstruction output; Constructing a normalized prior distribution : ； in, This is the normalized prior distribution. Indicates the total length of the sequence. This represents the original prior distribution value, i.e., the initial probability distribution, at time step v. This represents the original prior distribution value at time step u; Construct a series of consistency losses: ； in, Indicates a series of consistency losses. This is the normalized prior distribution. Indicates the Kullback-Leibler divergence. This represents the model prediction distribution at time step u; Constructing a priori stability loss: ； in, This represents the prior stability loss. Indicates to Smooth and enhance contrast. This represents the original prior distribution value at time step u; Define two loss functions with relative objectives: minimizing the objective and maximizing the objective. The minimizing objective is: ； The goal is to maximize: ； in, To adjust the parameters; During training, by minimizing the target loss and maximizing target loss Optimize and iterate the training process to obtain a well-trained anomaly detection model.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the unsupervised anomaly detection method for deep-sea manned submersibles based on multimodal feature fusion as described in any one of claims 1-7.

10. An unsupervised anomaly detection system for deep-sea manned submersibles based on multimodal feature fusion, characterized in that, include: The model building module is configured to: acquire multimodal signal data from deep-sea submersibles and perform preprocessing; An anomaly detection model is constructed. The preprocessed multimodal signal data is input into the anomaly detection model for training to obtain a trained anomaly detection model. The multimodal signal data of the deep-sea submersible is then input into the trained anomaly detection model to detect anomalies in the deep-sea submersible. The initial feature extraction module is configured to extract initial features from the preprocessed multimodal signal data to obtain local temporal features; The frequency domain enhancement module is configured to perform multi-scale frequency domain enhancement on local time-series characteristics to obtain frequency domain enhanced features. The frequency domain enhancement module includes multi-branch convolution operations and fusion operations. The multi-branch convolution operation extracts hidden frequency domain perturbation features, and the fusion operation fuses the frequency domain perturbation features. The long-term dependency modeling module is configured to: perform long-term dependency modeling on frequency domain enhanced features to obtain long-distance dependency features; The context enhancement module is configured to: use the fusion attention coding module to perform context enhancement on the frequency domain enhanced features to obtain coded features; The fusion attention encoding module models the correlation between features through serial attention and prior attention, enhancing the ability to capture local dependencies; The feature fusion module is configured to fuse frequency domain enhanced features, long-range dependency features, and coding features into multimodal features to obtain fused features; The reconstruction module is configured to: perform state modeling on the fused features, obtain global hidden state features, and reconstruct and restore the global hidden state features to obtain the reconstruction output; The optimization module is configured to use a joint loss function to optimize parameters and obtain a trained anomaly detection model through continuous iterative training.

Images